In the realm of meteorology, Time Series Forecasting Transformers provide the accurate forecasting of weather. The widely used zero input decoder’s initialization causes overfitting. To solve the problem of initializing decoder inputs, elaborate start tokens and mined trends from the encoder’s input but these are computationally expensive and not always effective for forecasting weather time series. We developed a Transformer with a fuzzy attention mechanism that is a transformed and decoupled Time Series Forecasting Transformer with less time and space complexity. As compared to fuzzy attention mechanisms with classical membership function, we developed a fuzzy attention mechanism that utilizes a context vector as fuzzy measures which interpretable insight into the forecasting process. By integrating fuzzy attention mechanisms and memory bottleneck step with transformer architectures, the proposed Transformer with a fuzzy attention mechanism integrates the strengths of Time Series Forecasting Transformers with the interpretability and flexibility of fuzzy logic, and provides a powerful framework for accurate and insightful weather forecasting. The simulation study results of the Transformer with a fuzzy attention mechanism shows its accuracy, robustness, and interpretability on future examples, as compared to classical Time Series Forecasting Transformer in weather forecasting.